We combine structural estimation with ideas from Machine Learning to estimate a model with information-based voluntary social distancing and state lockdowns to analyze the factors driving the effect of social distancing in mitigating COVID-19. The model allows us to estimate how contagious social interactions are by state and enables us to control for several unobservable, time-varying confounders such as asymptomatic transmission, sample selection in testing and quarantining, and time-varying fatality rates. We find that information-based voluntary social distancing has saved three times as many lives as lockdowns. Second, information policy effects are asymmetric: 'least informed' responses would have implied 240,000 more fatalities by June 2020 while 'most informed' responses would have saved 25,000 more lives. Third, our estimates suggest that contagion externalities from social interactions are large enough that a lockdown response could have been 25% less costly for the median state and still saved an equivalent number of lives.